AI Dating Profiles: How to Spot a Generated Match

A hospice nurse in her fifties, six weeks into the best conversation she has had in years, is asked for help with a customs fee on a shipment of equipment her match cannot clear from abroad. She has never seen his face move. Every photo is beautiful, and every one of them is slightly, unplaceably wrong in a way she cannot name.

AI dating profile detection works best as a sequence of cheap checks, not a single verdict from a tool. Read the photo set as a set, reverse image search everything, look for generation tells at full zoom, check provenance metadata when it survives, then test behavior. The conversation gives you better evidence than the pictures do.

How AI dating profile detection actually works

Three separate things get called fake, and they need different checks.

The oldest and still most common is a stolen real photo: a genuine person's pictures scraped from social media and reused. No AI involved, and no AI detector will flag it, because the image is real. Reverse image search is the tool that catches this one.

The second is a fully generated face, produced by a diffusion or GAN model. These have no origin anywhere on the internet, so reverse image search returns nothing, which is itself informative. Image detectors and visual tells are what work here.

The third is a real person with generated or heavily edited additions, or a real person whose face has been swapped into new scenes. This is the hardest case and the one where honest tools give hedged answers.

Now the part most articles skip. Detection on dating app photos is genuinely degraded, and you should know why before you trust a score. Every platform re-encodes images on upload, usually more than once. Bellingcat tested a leading image detector in September 2023 and found it missed seven of ten AI images after ordinary social-media-level compression. The signals detectors read live in fine pixel statistics, and compression is a machine for destroying exactly those. A screenshot of a screenshot is worse still.

So the honest framing is this: tools narrow your uncertainty, behavior resolves it. Someone who is who they say they are will get on a video call. Everything below is ordered by how much information it gives you per minute spent.

Step 1: Read the photo set as a set

Open every photo at once, side by side, and stop looking at faces for a moment.

Generated images are usually produced one at a time, which means consistency across a set is expensive to fake. Look for:

• Facial geometry drift. Ear position, the distance between eyes, the shape of a nostril. Small changes between photos that supposedly show the same person a month apart.
• Impossible continuity. A mole that migrates. A scar in one shot and not another. Teeth that change shape between smiles.
• Wardrobe and jewelry inconsistency. Earrings that change design in photos taken the same day, or a necklace whose chain does not connect.
• Uniform aesthetic. Every photo shot at the same distance, in the same soft light, with the same shallow background blur. Real camera rolls are messier than that.
• No unflattering photos. Real people have bad photos in their sets. Generated people rarely do.

The set-level view catches things a single-image inspection never will, and it costs about ninety seconds.

Step 2: Reverse image search every photo, not just one

Run each photo through a reverse image search. Do all of them, because scammers frequently mix sources.

Three outcomes and what each means:

Zero results is the ambiguous case, and it is where people over-read. A person who has never posted photos publicly will also return nothing. Do not treat an empty result as proof of anything. Treat it as the reason to keep going.

Step 3: Look for generation tells at full zoom

Zoom to full resolution and check the places models still get wrong. As of mid-2026 the crude tells of 2022 are mostly gone, but the failure modes have moved rather than disappeared.

• Hands and fingers, still, especially where fingers overlap an object or another hand.
• Text in the background. Signage, book spines, shop fronts. Generated text is the most reliable remaining tell, because models produce letter-shaped forms that spell nothing.
• Ears and earrings. Asymmetric ear structure, an earring on one side only, or jewelry that merges into the earlobe.
• Eyeglass frames that change thickness across the face or fail to align with the ears.
• Hair boundaries against a complex background, where individual strands dissolve into a soft edge.
• Reflections. Windows, sunglasses and mirrors that show something incompatible with the scene.
• Teeth count and alignment, which models render as an impression of teeth rather than a specific set.

Step 4: Check the file for provenance metadata

If you can save the image, it is worth checking whether any provenance record survived.

The C2PA Content Credentials standard embeds a signed history of how an image was made. OpenAI has attached it to image output since February 2024, and Adobe, Microsoft and Google's 2026 image models do the same. When a credential is present and says "generated," that is strong evidence. Google also watermarks Gemini image output with SynthID, though there is no public third-party API to verify it, so only Google's own tools can read that signal.

The catch, again: platforms strip this metadata routinely on upload. An absent credential means nothing at all. Present and confirming is useful; absent is normal.

Our AI image detector inspects embedded C2PA and EXIF data, reports a likely-generator attribution when the signal supports one, and gives an explicit confidence level rather than a bare percentage. It will also tell you plainly when a file has been compressed past the point where any conclusion is defensible, which is the answer you should want from a tool in this situation.

Step 5: Test the conversation, not just the pictures

Text is where the story falls apart, and it is free to examine.

Watch for the pattern rather than any single line. Romance fraud follows a recognizable arc: unusually fast emotional escalation, a profession that conveniently explains absence and wealth, a request to move off the platform to a private messaging app within days, and a gradual introduction of financial hardship framed as temporary and embarrassing.

Concrete tests you can run inside the conversation:

1. Ask a specific local question. If they say they live in Manchester, ask which tram line runs past their office. Generic answers to specific local questions are informative.
2. Reference something that did not happen. Mention "that restaurant you told me about last week" when they never did. A person corrects you. A script agrees enthusiastically.
3. Change register abruptly. Send something playful and idiomatic. Scripted and translated conversation gets brittle when the register moves.
4. Watch response timing against their claimed time zone. Consistent 3 a.m. enthusiasm in their stated city is a data point.

You can also scan the message text itself, and here you need calibration rather than confidence. Text detectors are unreliable on short samples, and dating messages are short by nature. Our free AI detector will scan up to 5,000 characters without a sign-up, but paste a long block of their messages rather than one line, and read the confidence label rather than the headline number. Why short samples produce noise, and how often detectors get things wrong in both directions, is covered in can AI detectors be wrong and the practical version in how to test text for AI in under five minutes.

Step 6: Ask for a live video call

This is the highest-information test available to you, and it takes two minutes.

Ask for an unscheduled call, not one arranged for tomorrow evening. Then, during it, ask them to turn their head fully to one side and to hold a hand up beside their face. Real-time face-swap tools exist and are getting better, but sharp profile angles and hands crossing the face remain their weak points as of mid-2026.

What you are really testing is willingness. In practice, most fraudulent accounts never reach the call. There is always a broken camera, a bad connection, a shift ending late. One postponement is life. Three is an answer.

If someone sends you a pre-recorded video instead, that is a different check. Our AI video detector analyzes footage frame by frame and returns a timeline showing where the synthetic-looking segments sit, which matters because manipulated clips are frequently real footage with a few altered seconds. A video check costs 25 credits, an image check 5, and a free account's 300 monthly credits covers sixty image checks, which is more than anyone needs for one profile.

Step 7: Report, block, and preserve the evidence

Order matters here, and most people get it wrong by confronting first.

Do not accuse them. The reliable outcome of "I think you're a bot" is that the account deletes itself, taking your evidence with it, and reappears tomorrow under a new name.

Screenshot everything first: the profile, the photos, the full conversation, any payment details or account numbers they gave you. Save the images as files where you can.

Report inside the app, using the specific category the platform offers for fake profiles or scams rather than a generic complaint. Platform review teams route by category.

If money changed hands, contact your bank or payment provider immediately and file with your national fraud reporting service. Speed matters more than completeness.

Then block. Not before, because blocking can cut your own access to the message history on some platforms.

One last thing worth saying plainly, because it is the part people carry longest: being deceived by a professionally operated fraud is not a failure of intelligence. These operations are staffed, scripted and iterated on thousands of targets. The nurse in the opening was not careless. She was targeted.

Nothing in this sequence produces certainty, and any product that tells you otherwise is selling something. Detection scores are evidence to weigh alongside behavior, and AI Detector 360 reports them with explicit confidence levels for exactly that reason. If you want the general version of these checks for any text, image or video you encounter, we cover it in is this AI. For a dating profile specifically, the video call is still the test that ends the question.